MiniMax model comparison: M2.7, M2.7-highspeed and M3

YiduoChan serves three MiniMax text models behind one OpenAI-compatible and one Anthropic-compatible endpoint. Picking between them is mostly a pricing and latency decision, and the rate card has three sharp edges worth understanding before you commit a workload.

The three MiniMax text models on YiduoChan — MiniMax-M2.7, MiniMax-M2.7-highspeed and MiniMax-M3 — are all reachable from the same base URL, with the same authentication and the same OpenAI-compatible request shapes. What separates them is the rate card, and that rate card has three edges you should design around: a 512K prompt-token price boundary on M3, an exact doubling of input and output on highspeed, and a hard 1,048,576-token ceiling on M3. The prices below are MiniMax's own published pay-as-you-go rates, which YiduoChan matches; every account currently sits in a group whose ratio is 0.95, so the amount actually billed is 5% below list.

The three text models at a glance

Two cells deserve a note. The M2.7 context window is not quoted here — see the M2.7 model page — and the M3 cache-write rate is not published, so if you cache heavily on M3, confirm what you were charged against your own usage records rather than assuming it matches M2.7. Model IDs are case-sensitive and must be sent exactly as written above.

M2.7 and M3 cost the same at or below 512K

This is the single most useful fact on the page. For any prompt at or below 512K tokens, M3 bills $0.30 input, $1.20 output and $0.06 cache read — identical to M2.7 on all three. Choosing M3 therefore costs you nothing at all until a prompt actually crosses the boundary. Above 512K and up to the 1M ceiling, input and output both double to $0.60 and $2.40, and cache read doubles to $0.12.

The practical rule: if your prompts might grow — a retrieval pipeline whose corpus is expanding, an agent that accumulates tool output, a document workflow whose inputs you do not control — set M3 as the default. You pay M2.7 rates in the common case and you keep headroom instead of hitting a wall. Reserve M2.7 for workloads whose prompt size you bound yourself and which will never approach 512K, where the extra headroom buys nothing.

What -highspeed actually buys

MiniMax-M2.7-highspeed is the same model quality at lower latency. On price it is exactly double M2.7 on input ($0.60 vs $0.30) and exactly double on output ($2.40 vs $1.20), while cache read is unchanged at $0.06. That asymmetry matters: the latency premium lands entirely on uncached input and on generated output, so a workload with a large, stable cached prefix pays a much smaller relative premium than an uncached one. The worked caching example below shows a case where switching to highspeed costs only 21.8% more rather than 100% more.

Treat highspeed as a latency purchase, not a capability purchase. If a slower response is merely mildly annoying, it is not worth the premium. If a response sits in a user-visible interactive loop — inline completion, a routing or classification hop in front of a bigger call, a voice-adjacent turn — it often is.

The 1M ceiling is hard

M3's 1,048,576-token context is a ceiling, not a soft target. There is no overflow tier above it and no graceful truncation to rely on: a prompt that exceeds it must be chunked, summarised or retrieved against before you send it. Note also that count_tokens is not available on the Anthropic-compatible endpoint, so any pre-flight size estimate has to come from your own tokenizer or a conservative character-based heuristic on your side.

Three worked monthly scenarios

All arithmetic below is at list price first, then multiplied by 0.95 for the current group discount. Rates are per 1,000,000 tokens.

1. Chat product on M2.7

120,000 requests per month, averaging 2,000 input tokens and 400 output tokens, no caching.

• Input: 120,000 × 2,000 = 240,000,000 tokens = 240M × $0.30 = $72.00
• Output: 120,000 × 400 = 48,000,000 tokens = 48M × $1.20 = $57.60
• List total: $72.00 + $57.60 = $129.60. After the 5% discount: $129.60 × 0.95 = $123.12

Run the same workload on M3 and the total is identical — $129.60 list, $123.12 discounted — because 2,000-token prompts sit far below 512K. Run it on highspeed and it doubles: 240M × $0.60 = $144.00 plus 48M × $2.40 = $115.20, giving $259.20 list and $246.24 discounted.

2. Long-context pipeline that crosses 512K

A document-analysis job on M3, 2,000 runs per month, 8,000 output tokens each. 1,200 runs have 400,000-token prompts (below the boundary) and 800 runs have 700,000-token prompts (above it).

• Below 512K, input: 1,200 × 400,000 = 480,000,000 = 480M × $0.30 = $144.00
• Below 512K, output: 1,200 × 8,000 = 9,600,000 = 9.6M × $1.20 = $11.52 (subtotal $155.52)
• Above 512K, input: 800 × 700,000 = 560,000,000 = 560M × $0.60 = $336.00
• Above 512K, output: 800 × 8,000 = 6,400,000 = 6.4M × $2.40 = $15.36 (subtotal $351.36)
• List total: $155.52 + $351.36 = $506.88. After the discount: $506.88 × 0.95 = $481.54

Forty percent of the runs carry 69.3% of the bill ($351.36 of $506.88). Re-chunking those 800 runs so each prompt lands at 500,000 tokens instead of 700,000 moves them back into the cheaper band: 800 × 500,000 = 400M × $0.30 = $120.00 input, plus 6.4M × $1.20 = $7.68 output, a subtotal of $127.68 instead of $351.36. New list total $155.52 + $127.68 = $283.20, or $269.04 after the discount — $223.68 of list savings from a chunking change alone. The cost calculator is the fastest way to test variants of this.

3. Latency-sensitive endpoint on highspeed

An interactive endpoint on MiniMax-M2.7-highspeed: 900,000 requests per month, 800 input tokens and 120 output tokens each, no caching.

• Input: 900,000 × 800 = 720,000,000 = 720M × $0.60 = $432.00
• Output: 900,000 × 120 = 108,000,000 = 108M × $2.40 = $259.20
• List total: $691.20. After the discount: $691.20 × 0.95 = $656.64

The same traffic on plain M2.7 is 720M × $0.30 = $216.00 plus 108M × $1.20 = $129.60, a list total of $345.60 and $328.32 discounted. The latency premium is $345.60 per month at list. Price that against what the latency is worth before shipping it.

How prompt caching changes the arithmetic

Cache read at $0.06 is one fifth of the $0.30 input rate, and cache write at $0.375 is 1.25× input. That ratio rewards one specific design: a long, byte-stable prefix at the front of every request, with the variable part appended at the end. Reordering, re-serialising or timestamping the prefix breaks the hit and you pay full input again.

A coding assistant on M2.7 with a 40,000-token stable prefix, 1,000 tokens of fresh turn text and 600 output tokens, across 50,000 requests per month. With no caching at all:

• Input: 50,000 × 41,000 = 2,050,000,000 = 2,050M × $0.30 = $615.00
• Output: 50,000 × 600 = 30,000,000 = 30M × $1.20 = $36.00
• List total $651.00, discounted $618.45

Now at a 90% cache hit rate, billing the 10% of prefixes that must be written at the cache-write rate:

• Cache reads: 45,000 × 40,000 = 1,800,000,000 = 1,800M × $0.06 = $108.00
• Cache writes: 5,000 × 40,000 = 200,000,000 = 200M × $0.375 = $75.00
• Fresh turn text: 50,000 × 1,000 = 50,000,000 = 50M × $0.30 = $15.00
• Output: 30M × $1.20 = $36.00
• List total $108.00 + $75.00 + $15.00 + $36.00 = $234.00, discounted $222.30

That is $417.00 of list savings, 64.1% off the uncached bill. Now run the cached version on highspeed: cache reads stay at $108.00 and cache writes stay at $75.00, only fresh input (50M × $0.60 = $30.00) and output (30M × $2.40 = $72.00) move. The list total becomes $285.00, or $270.75 discounted — just 21.8% above the M2.7 figure, not double. Verify the write-versus-input treatment against your own usage records before you budget on it.

Which model to set in Claude Code and Cursor

For Claude Code, point the Anthropic-compatible variables at YiduoChan. M3 is the sensible main model because agent sessions accumulate context, and highspeed is a good fit for the fast side-model slot:

export ANTHROPIC_BASE_URL=https://yiduochan.com
export ANTHROPIC_AUTH_TOKEN="<your key>"
export ANTHROPIC_MODEL=MiniMax-M3
export ANTHROPIC_DEFAULT_HAIKU_MODEL=MiniMax-M2.7-highspeed

# Use M2.7 instead when you bound prompt size yourself:
# export ANTHROPIC_MODEL=MiniMax-M2.7

For Cursor and anything else OpenAI-compatible, use the /v1 base URL and a bearer key. Streaming, tool calling and JSON mode all work through the standard shapes:

curl https://yiduochan.com/v1/chat/completions \
  -H "Authorization: Bearer $YIDUOCHAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "MiniMax-M2.7-highspeed",
        "messages": [{"role": "user", "content": "Reply with OK."}]
      }'

curl -s https://yiduochan.com/v1/models -H "Authorization: Bearer $YIDUOCHAN_API_KEY"

If you route per request, encode the two boundaries explicitly rather than leaving them implicit:

from openai import OpenAI

client = OpenAI(api_key="<your key>", base_url="https://yiduochan.com/v1")

M3_PRICE_BOUNDARY = 512 * 1024  # 524,288 tokens; above this M3 input/output/cache read double
M3_HARD_CEILING   = 1_048_576  # M3 will not accept more than this

def pick_model(prompt_tokens: int, latency_critical: bool = False) -> str:
    if prompt_tokens > M3_HARD_CEILING:
        raise ValueError("prompt exceeds the M3 1M ceiling & must be chunked")
    if latency_critical and prompt_tokens < HIGHSPEED_MAX_PROMPT:
        return "MiniMax-M2.7-highspeed"
    return "MiniMax-M3"

stream = client.chat.completions.create(
    model=pick_model(prompt_tokens=180_000),
    messages=[{"role": "user", "content": "Summarise the attached spec."}],
    stream=True,
)

Decision checklist

• Prompts might grow, or you do not control their size → MiniMax-M3. It costs exactly what M2.7 costs below 512K.
• Prompts are bounded and small, cost is the main concern → MiniMax-M2.7.
• Response latency is user-visible and worth roughly double on input and output → MiniMax-M2.7-highspeed.
• Prompts routinely exceed 512K → budget at $0.60 / $2.40 / $0.12, and check whether chunking below the boundary is cheaper than paying the upper band.
• Prompts may exceed 1,048,576 tokens → no model here covers that; chunk or retrieve first.
• A long prefix repeats across requests → freeze it byte-for-byte and let cache read at $0.06 do the work before you change models at all.

Billing notes

Billing is prepaid USD credits, pay-as-you-go, no subscription. The minimum top-up is $5, with presets at $5/$10/$20/$50/$100/$200/$500 or a custom amount, credits are valid for 12 months, and failed requests are never charged. There are no free credits and no trial, so plan a small first top-up while you validate the numbers above against real usage. The 5% group discount is the ratio currently applied to your account, not a permanent guarantee — check the pricing page for the live figures. Speech models are priced separately per input character; see speech-2.8 pricing. MiniMax-H3 video is temporarily unavailable on this platform. To start, see the MiniMax overview, the M3 model page, or create an account; questions go to [email protected].

FAQ

Sign up and get an API key See full model pricing